Possibilistic support vector machines for automatic phoneme recognition

Abstract

We propose, in this paper, new support vector machines (SVM) formulation that incorporates possibilistic weights based upon the geometric distribution of the phoneme's data set input to the recognition system. Those possibilistic weights are computed based on a possibilistic distance. Hence, we introduce a new formulation of the standard SVM incorporating the possibilitic weights (PossSVM). The experimental results show a greater performance of the proposed method than the existing SVM in the phoneme recognition task. Moreover, in this paper we tested several possibilistic distances in aim to find the most suitable with our data sets.